# Changelog

All notable changes to this project will be documented in this file.

## [1.1.9] - 2026-05-20
### Added
- **Gemini 3.5 Flash** (`gemini-3.5-flash`) — GA Vertex model with 1M input context, 65,535 max output tokens, reasoning, tool support, and $1.50/$9.00 per 1M token global pricing.
- **xAI Grok models** via Vertex MaaS: `grok-4.20-reasoning` and `grok-4.1-fast-reasoning`, including cache-read pricing.
- **Gemma 4 26B A4B IT** (`gemma-4-26b-a4b-it`) via Vertex MaaS.
- **Regional Claude pricing metadata** through optional `costRegional` model costs.

### Fixed
- Preserve Gemini 3/3.5 native thinking defaults when Pi reasoning is not explicitly requested by omitting `thinkingConfig` for those models.
- Apply a healthy default thinking budget for Gemini 2.5 models when Pi reasoning is not explicitly requested.
- Use regional Claude pricing when the resolved Vertex endpoint is non-global.

## [1.1.8] - 2026-05-06
### Fixed
- **Double `stream.end()` on the Anthropic path**: `streamAnthropic()` was calling `stream.end()` internally and then `streamMaaS()` was calling it again. Made `streamAnthropic()` lifecycle-neutral (pushes start/deltas/done but does not end the stream) so end() is called exactly once, matching the OpenAI-compat path. Idempotent in pi-ai today, but now correct by construction.
### Added
- **Tests for `streaming/maas.ts`** (`tests/streaming-maas.test.ts`): Anthropic happy path, tool-use stop reason, sync error path, OpenAI-compat model rewriting, and a regression test asserting `stream.end()` is called exactly once. Test count: 86 → 88.
### Changed
- **Removed dead branches in `convertToGeminiMessages`**: the `claude-` / `gpt-oss-` modelId checks (and the `requiresToolCallId` helper) only ran inside the Gemini streaming path, where modelId is always a Gemini apiId — they were unreachable in production. Three corresponding tests removed.
- **Lint cleanup**: tightened `any` usage in `index.ts`, `streaming/index.ts`, `streaming/gemini.ts`, and `utils.ts` (proper Tool / GeminiContent / discriminated-union types, exhaustive-check `never`). `noExplicitAny` is now disabled for `tests/**` (mock objects) and `streaming/maas.ts` (Anthropic message-shaping pipeline mixes intermediate shapes; cleanup tracked as a follow-up). `npm run check` is now a real signal: 53 warnings → 0.
- **Dropped `screenshot.png` from the npm tarball**: README now references the GitHub raw URL. Tarball size: ~730 kB → 24 kB (~30× smaller).
- **Added `.pi/` to biome ignore list** so internal task state isn't linted.

## [1.1.7] - 2026-05-06
### Added
- Claude Opus 4.7 model definition and README references using current Vertex metadata.
- Integration-style tests for `streamGemini()` with mocked `@google/genai` streaming responses.
- Additional Gemini conversion tests for valid Unicode, image tool results, and missing tool result synthesis.
### Fixed
- Preserve valid Unicode surrogate pairs while removing only unpaired surrogates before provider requests.
- Avoid double-counting Gemini cached input tokens as both uncached input and cache reads.
- Replay Gemini image tool results instead of silently dropping images.
- Insert synthetic Gemini tool results for missing tool call responses before replaying the next turn.
- Map unsupported Gemini 3 Pro `minimal` reasoning to `LOW` while preserving supported `MEDIUM` and `HIGH` levels.
- Use Gemini 2.5 Pro's lowest supported thinking budget when Pi reasoning is disabled instead of sending an invalid zero budget.
- Emit Gemini safety/blocked finishes as `error` stream events instead of `done` events with an invalid error stop reason.
- Update Claude 4.6 Vertex model metadata to current 128K output limits and current token pricing.

## [1.1.6] - 2026-05-05
### Added
- Comprehensive unit tests for `convertToGeminiMessages` (27 test cases covering user text, images, assistant text/thinking/tool calls, tool results, cross-provider signatures, and multi-turn conversations).
- Unit tests for `streamVertex` dispatch logic (gemini vs maas routing, unknown endpoint type errors).
### Fixed
- `streamAnthropic()` now calls `stream.end()` internally instead of relying on the caller, preventing potential stream hangs on early returns or mid-stream errors.
- Removed hardcoded `maxTokens / 2` halving in `streaming/gemini.ts` and `streaming/maas.ts`. Models now use their full advertised output capacity unless explicitly overridden via `options.maxTokens`.

## [1.1.5] - 2026-05-05
### Changed
- Forked to `lhl/pi-vertex` with standalone repository, CI, tests, and linting.
- Renamed package to `@lhl/pi-vertex`.
- Added Biome for linting and formatting.
- Added Vitest with coverage for unit tests (auth, config, utils, models).
- Added GitHub Actions CI workflow for type-check, lint, and test.
- Replaced placeholder `npm run check/build/clean` scripts with real implementations.

## [1.1.4] - 2026-03-30
### Fixed
- Removed error message override for `400 (no body)` responses from Vertex MaaS models. The original message now passes through to `isContextOverflow()` which already handles this pattern, enabling proper auto-compact instead of showing a raw error to the user.
- Use `zai` thinking format for `zai-org` publisher models (GLM-5). Previously using `openai` format which never sent `enable_thinking`, causing intermittent 400 errors from the ZAI API.

## [1.1.3] - 2026-03-26
### Fixed
- Hardened Claude-on-Vertex replay for mid-session model switching (tool ID normalization, tool result adjacency, thinking signature validation).
- Prevented Anthropic tool replay errors by inserting synthetic tool results when missing.

### Updated
- Claude 4.6 models use native Anthropic Vertex SDK streaming.
- Claude 4.6 context window updated to 1M.
- Model list order in the selector is now alphabetized by ID.

## [1.1.2] - 2026-03-24
### Changed
- Initial Claude 4.x support on Vertex.